event

Gemma Model Update

4 items · started 2026-09-28 · closed 2026-10-03

  1. I've shared a few deadlift demos in the past. This new demo includes some new models, accessed through the VLM Run Gateway: - SAM 3.1 to segment the barbell weight plate - ViTPose+ Large to measure the hip hinge angle, which serves as a ba…

  2. A worklog on megakernels for Gemma 4 on TPU v5e: 640 tok/s per user on the 26B-A4B MoE model versus 173 on an H100 with vLLM, why the dense 31B model gains only 1.2x, and when a megakernel is worth building.

  3. Gevva0: Calibrated Local Decision Gateway -00f2fe) []() A high-speed, local decision gateway running on Gemma 4 (llama.cpp / CUDA) that replaces flaky autoregressive JSON generation with direct token logit scoring, mathematical cyclic debi…

  4. When a language model receives two conflicting documents as input, how does it decide which one to prioritize? Does it rely on how the sources are framed or the presentation order of the documents?

← all threads